What is armada?
armada is a command-line tool that turns a repository into a team of eight opencode agents: an orchestrator, backend, frontend, QA, an adversarial reviewer, security, docs and a read-only architect. You and the orchestrator write a contract in armada/REQUIREMENTS.md with phases and success criteria; once you approve it, the agents build, test and review in a loop and finish by opening a pull request rather than merging locally.
Role boundaries are set through opencode's permission system: the orchestrator cannot edit source code, and the security, adversary and architect roles can only write their own review files. Every gate needs evidence such as a passing test run, a screenshot or a file and line citation, and loop state is written to disk so a killed session resumes. Each feature, called a voyage, runs in its own git worktree so several can run in parallel.
It is written in JavaScript, needs Node.js 22 or later and an authenticated opencode install, ships on npm as @rafamacalaba/armada, and is MIT licensed.
Is armada worth it?
armada's best decision is to enforce its maker and checker split with opencode's permissions instead of asking agents nicely. The orchestrator literally cannot edit code, and the reviewers can only write their own findings, so a builder cannot sign off on its own work. That is a sounder basis than most prompt packs, which rely on instructions an agent can ignore.
The evidence for it is thin. The README's main figure is that the fleet built armada's own session-state system in about 26 minutes for $0.18, catching and fixing 3 test failures along the way. That is one self-reported run by the author. The project has 105 stars, 12 forks and 2 open issues, and adoption outside the author is hard to judge.
It is also bound to opencode. The scaffold writes opencode.json and .opencode/agent files, the npm package lists opencode 1.18 or later as its peer, and voyages run in tmux sessions. Claude Code or Codex users get nothing from it without switching hosts. The armada models --discounts command points opencode at cheaper OpenRouter providers such as Novita, which the README says can save up to 20 times on some models; cheaper providers can also mean different quantisation and reliability, which the README does not discuss.
Activity has slowed. The repository was created on 1 August 2026, released v1.2.9 on 10 August and last received a commit on 28 August. That is well inside our 12-month line, but a two-month pause on a three-month-old project is worth weighing. Pullboard covers similar ground, with a spec, worktrees and a verifier agent, for any agent host and with far more published evidence.
Where it earns its place
A freelance developer already working in opencode
Hand a small feature to agents and get back a reviewed pull request
Run armada init in the repository, describe the feature to the orchestrator, approve the contract, and wait for the pull request with test runs, screenshots and the defect ledger attached. Model cost depends on the providers chosen; the author's own example cost $0.18.
A two-person startup on a tight model budget
Run several features in parallel without paying frontier-model prices
Start one voyage per feature, each in its own worktree, and route opencode to discounted OpenRouter providers with armada init --openrouter-provider. Check output quality on a cheap provider before relying on it.
Before you commit
- Setup
- npm install -g @rafamacalaba/armada on Node.js 22 or later with opencode installed and signed in; armada doctor checks the environment, then armada init scaffolds an existing repo or armada new starts one.
- Integrations
- opencode only (peer dependency 1.18 or later), git worktrees, tmux for parallel voyages, and OpenRouter provider routing for model choice; slash commands /voyage, /patrol, /fleet and /status inside opencode.
- Data and privacy
- Everything runs locally; contracts, state, defect ledgers, screenshots and E2E evidence are files under armada/ in your repo. Code goes to whichever model provider opencode is configured for.
- Cost at scale
- Free software; eight roles per feature multiply model calls, so cost tracks the provider and the number of review rounds. No published figures beyond the author's single $0.18 run.
- Limits
- Single maintainer, no commits since 28 August 2026, opencode-only, and no independent reports of use on large codebases.
What’s good
- Role limits enforced by opencode permissions, not prompts
- Adversary, security and QA reviewers are separate from builders
- Evidence required at each gate; state survives a killed session
- Parallel features in separate worktrees; ends in a pull request
- One runtime dependency besides opencode
What’s not
- Works only with opencode
- 105 stars; results are the author's own runs
- No commits since 28 August 2026
- Discount-provider routing trades price for unknown quality
Best for
opencode users who want a ready-made multi-role agent team with hard permission boundaries and are comfortable with a young, single-maintainer project.
armada alternatives
armada: common questions
- Is armada worth it?
- A well-reasoned agent team for opencode users that enforces roles by permission rather than by polite prompting, but small, quiet since August and tied to one host. We rate it 3.2 out of 5.
- How much does armada cost?
- Free under the MIT licence; you pay for the models opencode calls.
- What are the downsides of armada?
- Works only with opencode; 105 stars; results are the author's own runs; no commits since 28 August 2026; discount-provider routing trades price for unknown quality.
- Who is armada best for?
- opencode users who want a ready-made multi-role agent team with hard permission boundaries and are comfortable with a young, single-maintainer project.
- How much does armada cost for a team or in production?
- Free software; eight roles per feature multiply model calls, so cost tracks the provider and the number of review rounds. No published figures beyond the author's single $0.18 run.
- Is armada safe for company data?
- Everything runs locally; contracts, state, defect ledgers, screenshots and E2E evidence are files under armada/ in your repo. Code goes to whichever model provider opencode is configured for.
- What are the best armada alternatives?
- The closest alternatives we have reviewed are Pullboard, BMAD Method, Spec Kit and OpenCode.
Answers checked Nov 5, 2026.